The sequence
After Session 9, we will teach Sessions 11, 12 and 13, then move to the project. Session 10 is skipped. Earlier sessions remain part of the course. The session numbers follow the original course plan.
Pre-class reading
One bounded excerpt per session. For 11, read the specified parts of Shevlane et al.; for 12, the Introduction and Section 2 of Reuel et al.; for 13, the specified Key Concepts and evidence table in Olsson et al. Each session page gives the exact stopping point. Other papers are used in class or optional.
Session 11: can we trust the measurement?
Work through the evaluation taxonomy, elicitation and sandbagging, construct validity, and uncertainty and paired comparison. In the lab, reopen your Session 9 notebook and audit what its multilingual result does and does not show. Bring the completed Session 9 notebook to class.
Session 12: what does the measurement change?
Technical AI governance follows an eval through a threshold, a commitment and the question of independent verification. For the seminar, each group examines one frontier-safety framework and identifies a claim an outside auditor could or could not check. The African governance and compute-access questions give project routes beyond model experiments.
Session 13: what can we see inside?
Read why interpretability matters, then work through one induction-head circuit and its evidence limits. The session ends with a project scoping exercise.
Then the project
Use the project launch to choose a question, baseline and first-week result. The proposal is due at the end of the launch week. Robustness, evaluation and governance projects have taught starting points; interpretability projects require additional practical preparation and a narrower plan. The instructor will give calendar dates for the later clinic and final submissions.